How to provide image size for rectangular images?

[rahuja23]

Hello Guys! I have been training yolov5 on my custom dataset. I guess by default the model re-sizes the images to (640, 640). I trained my model on the re-sized images and have obtained a benchmark value for precision, recall and meAP scores. Now I would like to train the model with original image size (which is 960 x 528). Can someone help me with how can we give this image size during model training?

[glenn-jocher]

@rahuja23 python train.py --rect trains rectangular batches without mosaic.

[rahuja23]

ok and then I don' t have to specify the image size? or do I have to give the size as a tuple?

[glenn-jocher]

@rahuja23 you can define the rectangle long side length with --imgsz

[topherbuckley]

How can you specify both dimensions? --imgsz appears to only accept a single int according to this source. I tried a list and it failed:

  File "/media/HDD/git/beach-cleaning-robots/object-detection/.venv/lib/python3.10/site-packages/yolov5/train.py", line 720, in run
    main(opt)
  File "/media/HDD/git/beach-cleaning-robots/object-detection/.venv/lib/python3.10/site-packages/yolov5/train.py", line 615, in main
    train(opt.hyp, opt, device, callbacks)
  File "/media/HDD/git/beach-cleaning-robots/object-detection/.venv/lib/python3.10/site-packages/yolov5/train.py", line 208, in train
    train_loader, dataset = create_dataloader(train_path,
  File "/media/HDD/git/beach-cleaning-robots/object-detection/.venv/lib/python3.10/site-packages/yolov5/utils/dataloaders.py", line 124, in create_dataloader
    dataset = LoadImagesAndLabels(
  File "/media/HDD/git/beach-cleaning-robots/object-detection/.venv/lib/python3.10/site-packages/yolov5/utils/dataloaders.py", line 461, in __init__
    self.mosaic_border = [-img_size // 2, -img_size // 2]
TypeError: bad operand type for unary -: 'list'

Or does this somehow accommodate for varying short side lengths? If so, can you point to the source code that is doing this accommodation?

Edit:

Found what I was looking for here